A COMPACT, REPRODUCIBLE FIVE-GENE TRANSCRIPTOMIC CLASSIFIER DISTINGUISHES ESOPHAGEAL SQUAMOUS CELL CARCINOMA FROM NORMAL TISSUE: A MULTI-COHORT MACHINE-LEARNING STUDY

Abstract

Background. Esophageal squamous cell carcinoma (ESCC) is an aggressive malignancy with poor survival and late presentation. Robust, interpretable transcriptomic classifiers that generalise across independent cohorts remain scarce. We aimed to derive a compact, reproducible ESCC-vs-normal gene signature using interpretable machine-learning feature selection and to validate it in a fully independent cohort.

Methods. Three publicly available GEO microarray series were verified as ESCC tumour-vs-normal studies: GSE23400 (106 samples, HG-U133A), GSE77861 (14 samples, HG-U133_Plus_2), and GSE20347 (34 samples, HG-U133A_2). Expression values were already RMA-log2 normalised; probes were mapped to HGNC symbols and collapsed to a common universe of 13,039 genes. Discovery comprised GSE23400 + GSE77861 (120 samples; 60 tumour / 60 normal). GSE20347 (34 samples; 17/17) was held out as an independent validation set and never used for candidate selection, feature selection, hyperparameter tuning, or threshold optimisation. Candidate genes were defined by two strategies (Study-1 consensus DEGs and a broad variance filter) and reduced by LASSO, Random Forest (RF), and SVM-recursive feature elimination (SVM-RFE) within the discovery data only, using within-fold standardisation to prevent leakage. A consensus signature was assembled, compact panels (5/10/15/20 genes) were evaluated by repeated stratified cross-validation, and a Logistic Regression model was frozen and applied unchanged to GSE20347.

Results. Verification confirmed all three datasets represent ESCC (tumour vs normal). Two hundred ninety consensus DEGs and 2,000 high-variance genes formed the candidate pool (union = 2,001). LASSO selected 22 (λmin)/17 (λ1se) genes, RF 17 stable genes, and SVM-RFE 5 genes; the consensus contained 8 genes selected by ≥2 methods. Internal repeated cross-validation showed performance plateauing at the 5-gene panel (AUC = 0.998 ± 0.004). The frozen five-gene Logistic Regression signature (SIM2, RFC4, COL1A1, MMP1, CST1) achieved an independent-validation ROC-AUC of 1.00 (95% CI 1.00–1.00), PR-AUC 1.00, accuracy 0.941, sensitivity 0.882, and specificity 1.00. All five genes were individually significantly differentially expressed in GSE20347 (p < 0.001).

Conclusions. A compact five-gene expression signature generalises from discovery to a fully independent ESCC cohort with near-perfect discrimination. The signature is biologically plausible, interpretable, and suitable for future prospective validation.